Pooling in Graph Convolutional Neural Networks

Abstract

Graph convolutional neural networks (GCNNs) are a powerful extension of deep learning techniques to graphstructured data problems. We empirically evaluate several pooling methods for GCNNs, and combinations of those graph pooling methods with three different architectures: GCN, TAGCN, and GraphSAGE. We confirm that graph pooling, especially DiffPool, improves classification accuracy on popular graph classification datasets and find that, on average, TAGCN achieves comparable or better accuracy than GCN and GraphSAGE, particularly for datasets with larger and sparser graph structures.

Convolutional Neural Network for Malware Classification Based on API Call Sequence

“…This activation function rectifies the input values that have less than zero, by forcing them to be zero [23]. The following two layers flatten the generated feature map using mean-pooling [23,24] on segments of length two within the generated feature map.…”

“…Pooling is used to gradually decrease the dimensions of the attribute representation. Hence, it provides low computational cost by shrinking memory cost and the number of parameters [23,24]. Mean-pooling calculates the average of each feature values of the feature map [24].…”

“…Hence, it provides low computational cost by shrinking memory cost and the number of parameters [23,24]. Mean-pooling calculates the average of each feature values of the feature map [24]. The next layer is a standard feed-forward layer in a deep learning architecture which transforms the generated feature vector using ReLU activation [22], shown in Eq.…”

Untitled

“…Pooling is used to gradually decrease the dimensions of the attribute representation. Hence, it provides low computational cost by shrinking memory cost and the number of parameters [28][29].Average pooling and max-pooling are the most popular pooling techniques. While average pooling calculates the average of each feature values of the feature map [29], max pooling instead returns the maximum value of the feature values.…”

“…This layer uses 64 filters with ReLU (Rectified linear unit function) activation function [27], which has an almost linear function and therefore retains the properties of linear models. These properties provide ease to optimize with gradient descent methods.The following two layers flatten the generated feature map using average-pooling [28][29] on segments of length two within the generated feature map.The next layer is a standard feed-forward layer in a deep learning architecture which transforms the generated feature vector using ReLU activation [27]. The final layer then uses softmax activation [27] to generate a vector.…”